LC Sigmas (BETA)

Schedulers are the knob everyone leaves on default. That was fine when the answer was always Karras. On flow-matching models Karras and exponential went from safe to actively harmful - a straight trajectory has nothing for an aggressive noise schedule to fix - and the real shortlist now lives in a third-party pack, which is why workflows naming beta57 won't even load without RES4LYF installed. LC Sigmas is the attempt to put all of that in one node, with the maths built in rather than borrowed.

What it is

A scheduler node. It builds a sigma schedule from your model's own noise range and shift, and it handles one-pass and two-pass sampling in the same place: the curve, the step swap to a second model, image-to-image denoise, what the second pass does at the handover, and the noise and sampler for each pass.

sigma_curve_presets is the shape of the schedule:

• base - a plain ComfyUI scheduler chosen in base_scheduler (simple, sgm_uniform, karras, exponential, beta, beta57, bong_tangent, kl_optimal and friends).
• beta - Comfy's beta schedule with your own beta_alpha and beta_beta.
• flowmatch - the FlowMatch Euler Discrete curve, with flowmatch_shift, flowmatch_dynamic and flowmatch_terminal.
• hyperbolic and gaussian - the base curve bent by sinh/cosh/tanh and friends, or by a bell curve (gaussian_operation pdf/cdf/inverse_cdf/transform/modulate). These are the same maths as RES4LYF's Sigmas Hyperbolic and Sigmas Gaussian nodes.
• saved: … - your own curves, saved from here or from LC Sigma Curve, living in ComfyUI/user/LC123/sigma_curves/ where an update can't eat them.

And the graph on the node isn't decorative. Drag a point and the curve bends: edit_mode smooth pulls neighbouring steps along, spike moves only that point, and smooth_radius says how many neighbours come along. reset curve clears the edit. save curve arms a save - you still have to queue the workflow for it to happen.

The two-pass half

step_swap is the step where model_2 takes over (0 = one pass). The second pass starts at exactly the noise level the first stopped at, and the node lines up two different architectures by comparing noise as a share of noise in the image, which is what makes an SDXL-esque model hand off cleanly to a flow model. scheduler_shape_2 gives the second pass its own curve, or follows the first.

low_pass decides what the handover means:

• continue noise schedule - the second model takes the leftover noise and finishes the job. One schedule, two samplers.
• renoise - the first pass finishes the picture, fresh noise goes back up to the swap level, and the second pass refines from there. That's a detailer, and it's the setting people want when pairing a base model with a different refiner.

denoise (with denoise_steps set to full schedule or at step swap) covers image-to-image. first_pass_resample spreads the first pass over more steps along the same path without moving where it starts or stops. sigma_shift bends the schedule without changing the step count - use model default, by image size (Flux's rule; needs image_size_latent wired), or custom with custom_shift.

Outputs, and the bit to actually read

Seven of them: sigmas_high and sigmas_low (the two passes), noise_high and noise_low (NOISE objects, with noise_low derived from seed + 1), sigmas_full for the unsplit schedule, sampler - the enum choice handed out as a real SAMPLER for SamplerCustomAdvanced - and info, a plain-language string describing what the schedule does. Wire sigmas_high, noise_high and sampler into SamplerCustomAdvanced and you're running.

The sampler dropdown lists 44 options and pulls in RES4LYF's samplers when it's installed; a ClownSampler Selector can be wired in too. More usefully, the line under the graph recommends a pairing for the model you plugged in, from the pack's own testing: Krea 2 → euler + beta57; Z-Image → uni_pc + simple; Flux 2 → uni_pc/simple or dpmpp_2m/sgm_uniform; SDXL, Illustrious and Pony → dpmpp_2m + karras.

Worth knowing where that recommendation breaks: with a large sigma shift (say ModelSamplingAuraFlow at 4 in front of Krea 2), beta57 crams the last steps into very low noise, and a Turbo-based model starts drawing grain there - which Detail Daemon then turns into crunchy skin. At high shift, simple stays photographic. The node accounts for this, but now you know why.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/lonecatone23/ComfyUI_LC123_nodes.git

Restart, or search ComfyUI LC123 Nodes in ComfyUI Manager. No requirements.txt, no pip. The node is still marked BETA, and the live curve editor is a classic-canvas feature: in Nodes 2.0 you get a picture of the curve after the run instead.

Where it bites

Two practical traps. A high shift combined with a low-end-crowding schedule - crunchy skin, inked-looking eyes - so change the shape before you blame the sampler. And expecting renoise and continue noise schedule to behave alike: one hands over, the other restarts. If you're on an SDXL finetune, none of this buys you anything - dpmpp_2m with Karras is still correct there.